Separating Actor-View from Speaker-View Opinion Expressions using Linguistic Features

Michael Wiegand, Marc Schulder, Josef RuppenhoferNAACL 2016

TL;DR

This paper tackles the nuanced task of distinguishing between opinions expressed by the actor of an event and opinions expressed by the speaker of an utterance, focusing on individual words. They introduce a rich set of linguistic features and demonstrate that Markov Logic Networks (MLNs) effectively leverage these features, especially when incorporating global constraints, to achieve state-of-the-art performance in this word-level categorization.

What Problem Does This Paper Address?

Sentiment analysis typically focuses on identifying whether an expression is subjective and what its polarity (positive/negative) is. However, a deeper understanding requires knowing whose opinion is being expressed. This paper addresses the challenge of differentiating between two types of opinion expression: actor views, where the sentiment is directed towards entities involved in the event described by an opinion word (e.g., "Party members were disappointed actor at the election outcome"), and speaker views, where the sentiment is expressed by the speaker of the utterance, often implicitly (e.g., "Sarah excelled speaker in virtually every subject"). This distinction is crucial for tasks like opinion holder and target extraction, as the same word can function differently depending on the view it conveys. While previous work has explored this in context, this paper focuses on a more granular, word-level categorization, aiming to build a lexicon of opinion words with their inherent view type.

Methods

The authors propose a supervised classification approach using Markov Logic Networks (MLNs). MLNs are powerful because they combine first-order logic with probabilities, allowing for the incorporation of both local features and global constraints.

They design a comprehensive set of linguistic features, categorized as follows:

  • WordNet Glosses (GLOSS): Bag-of-words features derived from the definitions of words in WordNet. This serves as a baseline.
  • Lexicographer Files (LEX): Coarse-grained semantic categories from WordNet, requiring less labeled data.
  • FrameNet (FN): Semantic frames from FrameNet, assuming different frames are associated with different sentiment views.
  • Subcategorization Frames (SUB): Syntactic information about how opinion words are used in sentences, particularly the arguments they take.
  • Morphological Information (MORPH): Derivational affixes (prefixes and suffixes) that can indicate sentiment.
  • Context Patterns (PATT): Patterns identified in previous work for actor-view and speaker-view verbs, extended to nouns and adjectives.
  • Polarity Information (POLAR): Pre-assigned polarity (positive, negative, neutral) from the Subjectivity Lexicon.

Crucially, they also incorporate global constraints into the MLNs. These constraints enforce consistency, stating that if two opinion words are similar (either via Word2Vec embeddings, Lin similarity, or morphological relatedness), they should have the same sentiment view.

Datasets

The primary dataset used is a manually annotated lexicon of verbs, nouns, and adjectives derived from the Subjectivity Lexicon (Wilson et al., 2005). This dataset is an extension of previous work by Wiegand and Ruppenhofer (2015). For context patterns, they utilized the North American News Text Corpus (LDC95T21). They also reference the MPQA corpus for evaluating the correlation between their prior word-level annotations and contextual sentiment views.

Key Insights and Ideas

  • Word-level categorization is feasible and valuable: The paper demonstrates that opinion words can be reliably categorized into actor and speaker views even without full sentence context, which is crucial for building comprehensive lexicons.
  • Linguistic features are highly predictive: A diverse set of features, especially subcategorization frames, FrameNet information, and morphological cues, prove to be very effective in distinguishing between actor and speaker views.
  • Global constraints significantly boost performance: Incorporating consistency constraints between similar words within MLNs leads to substantial improvements, particularly for nouns and adjectives, highlighting the benefit of modeling relationships between instances.
  • Actor-view nouns can be tricky: The paper notes that actor-view nouns are more prone to confusion with speaker views, especially when they appear in longer or more complex subjective phrases.
  • Prior labels correlate with contextual views: The word-level annotations largely align with contextual annotations in the MPQA corpus, validating the approach and suggesting that the compiled lexicon can be useful for downstream tasks.

Key Results

The proposed MLN-based approach with a comprehensive feature set and global constraints significantly outperforms a graph-based approach and SVM baselines. For instance, the best performing classifier (MLN with global constraints and graph features) achieved an F1 score of 76.9% for nouns, a notable improvement over the baseline SVM's 72.4%. The paper also shows that features from most groups contribute complementary information, as removing any single group leads to a statistically significant performance drop.

Limitations and Future Work

  • Limited training data: The authors focus on settings with limited labeled data, which is practical for lexicon induction but might limit the performance of more complex models. They show their approach is effective even with only 20% of the data.
  • Actor-view vs. Speaker-view distinction for nouns/adjectives: While verbs can be further categorized into agent/patient views, this distinction is not directly applicable to nouns and adjectives, leading to a merged "actor view" category for these parts of speech.
  • Contextual mismatches: Despite strong correlations, mismatches between prior and contextual labels exist, particularly for actor-view words in speaker-view contexts. This is partly attributed to the MPQA annotation focusing on salient opinions and larger spans.
  • Future work: The authors suggest exploring more sophisticated ways to model the relationship between salience and contextual sentiment views, and further investigating the nuances of actor-view nouns in speaker-view contexts. They also hint at the potential for their approach to be applied to larger, open-domain sentiment lexicons.

Summary updated

BibTeX
@inproceedings{wiegand2016separating,
  author = {Michael Wiegand and Marc Schulder and Josef Ruppenhofer},
  title = {{Separating Actor-View from Speaker-View Opinion Expressions using Linguistic Features}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
}
Figure